Systematic Parameterized Description of Pro-Forms in the Prague Dependency Treebank 2.0∗

Systematic Parameterized Description of Pro-Forms in the Prague Dependency Treebank 2.0∗

Systematic Parameterized Description of Pro-forms in the Prague Dependency Treebank 2.0¤ Magda ŠevcíkovᡠRazímová, Zdenekˇ Žabokrtský Institute of Formal and Applied Linguistics Charles University, Prague E-mail: {razimova,zabokrtsky}@ufal.mff.cuni.cz 1 Introduction A pro-form is a word that is used to replace or substitute other words, phrases, clauses, or sentences etc. Besides pronouns one can also distinguish pro-adjectives, pro-numerals, pro-adverbs, and pro-verbs.1 Pro-forms are related to a wide range of linguistic phenomena, from word- formative principles, through negation and quantification, to anaphoric and deictic functions. As it was recognized a long time ago (among others in Montague gram- mar), pro-forms are extremely important for studying natural language semantics, even if they constitute only a closed class. Recently, a lot of work has been invested into developing large data resources for exploring natural language semantics, e.g. in the fields of predicate-argument structures or lexical databases. However, the treatment of pro-forms and related phenomena receives only a relatively minor attention in this data-dominated era (perhaps with the only exception of data for anaphora resolution, mostly limited to personal and demonstrative pronouns). Even if tag sets used in various corpora and treebanks clearly indicate some differentiation within the set of pro-forms (e.g. ¤The research reported in this paper was supported by the projects 1ET101120503, GA-UK 352/2005 and GD201/05/H014. 1The well known difficulties with the heterogeneity of the criteria for delimitating the ‘traditional’ parts of speech lead to terminological confusion here: pronouns are often considered to span not only pro-nouns, but also some of the other pro-forms, especially pro-adjectives and also pro-adverbs, often denoted as pronominal adjectives and pronominal adverbs. The term pronominal nouns is used less frequently, perhaps because it sounds pleonastic. The term pronominal verb is mostly used to denote not a pro-verb, but a verb accompanied with a reflexive particle. wh-words in English), they do not allow us to directly observe and employ many semantically significant analogies present in the pro-form systems.2 This concerns, e.g., the fact that nobody, never, and nowhere share certain semantic feature in their meanings, as well as everybody, always, and everywhere do, and that the two features are mutually exclusive. Moreover, the present tag sets in some cases do not distinguish expressions which have the same surface shape but which significantly differ in semantic or pragmatic aspects. For instance, personal pronouns used in formal (esteemed) speaking may be homonymous with other pronouns (e.g. third person plural in German or second person plural in Czech). Similarly, interrogative and relative pronouns are known to be ambiguous in many Indo-European languages and they also usually obtain the same POS tag, although the difference between them would become crucial when constructing e.g. a dialog system.3 In this paper we present a formal linguistic system for the annotation of pro- forms which has been developed and implemented in the framework of the tec- togrammatical layer of the Prague Dependency Treebank 2.0. The main motiva- tion of our approach is the following: if there is a semantically relevant regularity within a certain subset of pro-forms, then it is more useful – at least from the view- point of treebank users interested in natural language semantics, in conversions into logical forms etc. – if such information is available in the treebank in an explicit, machine-tractable form. In this case, the semantic features originally present in the word form (given its context) are extracted and stored as values of inner parameters of tectogrammatical nodes corresponding to the given word form. Metaphorically, this can be seen as snatching pieces from the lexical space and reshaping them into multidimensional orthogonal blocks in which the semantics of each element can be derived from the semantics of its coordinates in an entirely compositional fashion. Of course, the question of regularities in the pro-form systems is by far not new; various attempts at systematizing (at least certain subsets of) pro-forms can be found e.g. in [8], [7], [1], or in Wikipedia.4 What we believe is new here is that the presented system is explicit and implementable, incorporated into the elaborated system of deep-syntactic analysis, and at the same time, applied (and verified) on large data. 2Besides the set of morphological tags used in the Prague Dependency Treebank, we have studied also rules for tagging pro-forms in Penn Treebank ([10]), Tiger Treebank ([6]), MULTEXT-East projects ([4]), and BulTreeBank ([13]) from the perspective of pro-forms. The last one seems to be the most developed in this aspect. 3However, although the distinction between ambiguous relative and interrogative pro-forms is not explicitly marked e.g. in the Penn Treebank, it could be derived from the shape of the phrase- structure annotation with a high precision. 4http://en.wikipedia.org/wiki/Pro-forms 2 PDT 2.0 in a Nutshell In the Prague Dependency Treebank annotation scenario, based on the theoretical framework of Praguian Functional Generative Description ([12]), three layers of annotation are added to Czech sentences (see Figure 1):5 Morphological layer (m-layer), on which each token in each sentence of the source texts is lemmatized and tagged with a positional POS-tag.6 Analytical layer (a-layer), on which a sentence is represented as a rooted or- dered tree with labeled nodes and edges, corresponding to the surface-syntactic relations; each a-layer node corresponds to exactly one m-layer token. Tectogrammatical layer (t-layer), on which a tree structure of a sentence is la- beled with tectogrammatical lemmas (often different from the morphological ones) and dependency relations (semantic roles, functors) and enriched with valency an- notation, annotation of coreference, topic-focus annotation and annotation of se- mantically relevant grammatical meanings (grammatemes) and related attributes for node classification such as sempos (semantic part of speech). Annotations at all tree layers (m-layer, a-layer, and t-layer) are part of PDT 2.0. PDT 2.0 data consist of 7,110 manually annotated textual documents, containing altogether 115,844 sentences with 1,957,247 tokens (word forms and punctuation marks). All these documents are annotated at the m-layer, 75 % of them are an- notated at the a-layer. 59 % of the a-layer data are annotated also at the t-layer (i.e. 45 % of the m-layer data; 3,165 documents, 49,431 sentences, 833,195 to- kens). The CD-ROM including the final annotation of PDT 2.0 data, a detailed documentation as well as software tools has been publicly released by Linguistic Data Consortium in 2006 ([3]).7 3 Pro-forms in the PDT 2.0 At the m-layer, pronouns, pronominal adverbs, and pronominal numerals are treated, as usual, separately. The part-of-speech information is encoded in the first of the 15 tag positions: the upper case letter P stands for pronouns, D for adverbs and C for numerals. The part-of-speech information is further specified at the second tag position (see the left column in Figure 2). Neither pronominal adverbs nor 5See http://ufal.mff.cuni.cz/pdt2.0/ for a detailed documentation and sample data of PDT 2.0. 6Technically, there is also one more layer called w-layer (word layer) ‘below’ the m-layer; on this lowest layer the original raw text is only segmented into documents, paragraphs and tokens, and all these units are enriched with identifiers. 7The previous version of the treebank, PDT 1.0, was smaller and contained only m-layer and a-layer annotation (see [2]). Figure 1: PDT 2.0 annotation layers and the layer interlinking illustrated (in a highly simplified fashion) on the sentence Byl by šel do lesa ([He] would have gone into [a] forest). pronominal numerals are delimited as special subclasses of the respective parts of speech. While m-layer annotation8 was assigned only to pro-forms that are present in the surface shape of the sentence, at the t-layer we have to deal also with pro- forms that do not correspond to any word in the outer shape of the sentence – in the sequel, we call them restored nodes. If a restored node stands e.g. for a pro- dropped subject (as in Figure 1) that is not present in the surface sentence, it is considered to be a personal pronoun at the t-layer.9 At the t-layer, we have developed two different annotation schemes for pro- forms. The first scheme, which we present in Section 3.1, has been suggested for personal pronouns, taking into account the special character of these pronouns (e.g. they have a strictly deictic function, lacking a real lexical meaning). The second 8Note that the a-layer annotation does not add any information specifically related to pro-forms. 9Not only the new node is added to the structure, but also the values of its grammatical categories such as gender or number are reconstructed in the PDT 2.0 data (using e.g. subject-verb agreement or coreference relations). Figure 2: Rearrangement of pro-forms during the transition from m-layer to t-layer (note that arrows corresponding to other than pronominal numerals and adverbs are not displayed in the figure). At the m-layer, each pro-form is represented by its word form, morphological lemma and positional tag (values of the first one or two positions are specified in the entries in the left column). At the t-layer, pro-forms are represented as labels of t-nodes; for each of them, the attribute sempos (de- tailed semantic part of speech, see [11] for the explanation of the two-level t-node type hierarchy) specifies which other attributes are to be filled (besides tectogram- matical lemma, which is always obligatory).

View Full Text

Details

  • File Type
    pdf
  • Upload Time
    -
  • Content Languages
    English
  • Upload User
    Anonymous/Not logged-in
  • File Pages
    12 Page
  • File Size
    -

Download

Channel Download Status
Express Download Enable

Copyright

We respect the copyrights and intellectual property rights of all users. All uploaded documents are either original works of the uploader or authorized works of the rightful owners.

  • Not to be reproduced or distributed without explicit permission.
  • Not used for commercial purposes outside of approved use cases.
  • Not used to infringe on the rights of the original creators.
  • If you believe any content infringes your copyright, please contact us immediately.

Support

For help with questions, suggestions, or problems, please contact us